Sensing With Less Effort: Knowledge Transfer for Wireless Sensing by Unifying Feature Embeddings

Wireless sensing is an enabling technology that equips wireless devices with powerful sensing capabilities by analyzing the effect of object targets on surrounding wireless signals. With excellent feature extraction and analysis capabilities, deep neural networks have long become a promising tool for implementing wireless sensing. However, because wireless signals are not human-interpretable, the practical deployment of wireless sensing systems is hindered by the labor-intensive effort of collecting and labeling training samples. This article attempts to provide a comprehensive view of the field of wireless sensing and proposes a novel knowledge transfer method for accomplishing wireless sensing tasks with less effort. Specifically, we first introduce the framework of a deep neural network-based wireless sensing system and review the classical wireless sensing methods with less effort. Then, we highlight our proposed knowledge transfer method by unifying feature embeddings, which will utilize a small number of synchronously collected videos to realize the alignment of wireless and video features in a unified feature space, transfer the knowledge from the video dataset to the wireless sensing network, and thus reduce the effort of wireless sample collection and labeling in the training phase. We also evaluate our proposed method using in-home daily status sensing as a case study. Finally, we discuss existing challenges and future research directions.